SmallInitEmb - LayerNorm(SmallInit(Embedding)) in a Transformer to improve convergence

Overview

SmallInitEmb

LayerNorm(SmallInit(Embedding)) in a Transformer

I find that when training a transformer, the embedding matrix moves slowly, hence it's difficult for the model to jump out of the initial noisy embedding.

(initial embedding)
[[-0.0073  0.0062 -0.0261 ...  0.0086  0.0107 -0.008 ] ... ]
 (after 1 step, the directions of the embedding vectors are not moved much because the numbers change by ~LR = ~4e-4)
[[-0.0069  0.0066 -0.0265 ...  0.009   0.0111 -0.0084] ... ]

So I propose initializing the embedding matrix to tiny values, and put another LayerNorm after it (before all the SA & FFN layers):

if isinstance(module, (nn.Embedding)):
    nn.init.uniform_(module.weight, a=-1e-4, b=1e-4) # SmallInit(Emb)
...
if self.config.USE_SMALL_EMB and self.layer_id == 0:
    x = self.lnPre(x) # LN(SmallInit(Emb))
x = x + self.att(self.ln1(x))
x = x + self.ffn(self.ln2(x))

And then you get improved convergence (especially for BPE models) because the model can quickly jump out of the tiny initial embedding (small changes after 1 step -> significant changes of directions -> significant changes after LayerNorm).

Loss curve comparison: https://wandb.ai/blinkdl/SmallEmbTest

(the gap between LayerNorm(SmallEmb)) and baseline persists after more training)

Moreover, you can directly train PostLN models without warmup with SmallInit(Emb)

if isinstance(module, (nn.Embedding)):
    nn.init.uniform_(module.weight, a=-1e-4, b=1e-4) # SmallInit(Emb)
...
x = self.ln1(x) # this plays the same role as the lnPre in the above PreLN code
x = x + self.att(x)
x = self.ln2(x)
x = x + self.ffn(x)
(note you shall have another LN after the final ffn)
Owner
PENG Bo
http://zhihu.com/people/bopengbopeng
PENG Bo
All materials of Cassandra Event, Udyam'22

Cassandra 2022 Workspace Workshop Materials Workshop-1 Workshop-2 Workshop-3 Workshop-4 Assignments Assignment-1 Assignment-2 Assignment-3 Resources P

36 Dec 31, 2022
NeurIPS 2021 Datasets and Benchmarks Track

AP-10K: A Benchmark for Animal Pose Estimation in the Wild Introduction | Updates | Overview | Download | Training Code | Key Questions | License Intr

AP-10K 82 Dec 11, 2022
This is the official code of L2G, Unrolling and Recurrent Unrolling in Learning to Learn Graph Topologies.

Learning to Learn Graph Topologies This is the official code of L2G, Unrolling and Recurrent Unrolling in Learning to Learn Graph Topologies. Requirem

Stacy X PU 16 Dec 09, 2022
Misc YOLOL scripts for use in the Starbase space sandbox videogame

starbase-misc Misc YOLOL scripts for use in the Starbase space sandbox videogame. Each directory contains standalone YOLOL scripts. They don't really

4 Oct 17, 2021
This repository contains the re-implementation of our paper deSpeckNet: Generalizing Deep Learning Based SAR Image Despeckling

deSpeckNet-TF-GEE This repository contains the re-implementation of our paper deSpeckNet: Generalizing Deep Learning Based SAR Image Despeckling publi

Adugna Mullissa 16 Sep 07, 2022
A multi-functional library for full-stack Deep Learning. Simplifies Model Building, API development, and Model Deployment.

chitra What is chitra? chitra (चित्र) is a multi-functional library for full-stack Deep Learning. It simplifies Model Building, API development, and M

Aniket Maurya 210 Dec 21, 2022
This is the code repository for the paper "Identification of the Generalized Condorcet Winner in Multi-dueling Bandits" (NeurIPS 2021).

Code Repository for the Paper "Identification of the Generalized Condorcet Winner in Multi-dueling Bandits" (To appear in: Proceedings of NeurIPS20

1 Oct 03, 2022
The pytorch implementation of the paper "text-guided neural image inpainting" at MM'2020

TDANet: Text-Guided Neural Image Inpainting, MM'2020 (Oral) MM | ArXiv This repository implements the paper "Text-Guided Neural Image Inpainting" by L

LisaiZhang 75 Dec 22, 2022
Python package for visualizing the loss landscape of parameterized quantum algorithms.

orqviz A Python package for easily visualizing the loss landscape of Variational Quantum Algorithms by Zapata Computing Inc. orqviz provides a collect

Zapata Computing, Inc. 75 Dec 30, 2022
Code for our paper "Multi-scale Guided Attention for Medical Image Segmentation"

Medical Image Segmentation with Guided Attention This repository contains the code of our paper: "'Multi-scale self-guided attention for medical image

Ashish Sinha 394 Dec 28, 2022
Official PyTorch Implementation of SSMix (Findings of ACL 2021)

SSMix: Saliency-based Span Mixup for Text Classification (Findings of ACL 2021) Official PyTorch Implementation of SSMix | Paper Abstract Data augment

Clova AI Research 52 Dec 27, 2022
[ICML 2020] DrRepair: Learning to Repair Programs from Error Messages

DrRepair: Learning to Repair Programs from Error Messages This repo provides the source code & data of our paper: Graph-based, Self-Supervised Program

Michihiro Yasunaga 155 Jan 08, 2023
Official Pytorch Implementation of GraphiT

GraphiT: Encoding Graph Structure in Transformers This repository implements GraphiT, described in the following paper: Grégoire Mialon*, Dexiong Chen

Inria Thoth 80 Nov 27, 2022
💊 A 3D Generative Model for Structure-Based Drug Design (NeurIPS 2021)

A 3D Generative Model for Structure-Based Drug Design Coming soon... Citation @inproceedings{luo2021sbdd, title={A 3D Generative Model for Structu

Shitong Luo 118 Jan 05, 2023
Source code of our BMVC 2021 paper: AniFormer: Data-driven 3D Animation with Transformer

AniFormer This is the PyTorch implementation of our BMVC 2021 paper AniFormer: Data-driven 3D Animation with Transformer. Haoyu Chen, Hao Tang, Nicu S

24 Nov 02, 2022
Multiple-Object Tracking with Transformer

TransTrack: Multiple-Object Tracking with Transformer Introduction TransTrack: Multiple-Object Tracking with Transformer Models Training data Training

Peize Sun 537 Jan 04, 2023
Collection of machine learning related notebooks to share.

ML_Notebooks Collection of machine learning related notebooks to share. Notebooks GAN_distributed_training.ipynb In this Notebook, TensorFlow's tutori

Sascha Kirch 14 Dec 22, 2022
Guiding evolutionary strategies by (inaccurate) differentiable robot simulators @ NeurIPS, 4th Robot Learning Workshop

Guiding Evolutionary Strategies by Differentiable Robot Simulators In recent years, Evolutionary Strategies were actively explored in robotic tasks fo

Vladislav Kurenkov 4 Dec 14, 2021
Playing around with FastAPI and streamlit to create a YoloV5 object detector

FastAPI-Streamlit-based-YoloV5-detector Playing around with FastAPI and streamlit to create a YoloV5 object detector It turns out that a User Interfac

2 Jan 20, 2022